Research On Socially Compliant Navigation Of Robot In Dynamic Environment

Robot navigation technology gives the robot the ability to move freely in space,so that the robot’s working scene is no longer limited to fixed points,and can shuttle more flexibly through different scenes to complete a variety of tasks.The existing robot navigation technology can work well in fixed,small-scale and feature-rich environment,such as short-distance transportation in unmanned factory,guidance in small fixed area and so on.However,mobile robots are rarely seen in the large,dynamic and non-structural environments with frequent human-robot interaction,such as shopping malls and airports.Most of the robots actually running in such environments are rigid and conservative,and need to brake or re-planning frequently to ensure their own safety.The main reason are that the dynamic large-scale environment where multi-agent coexistence brings great challenges on robot navigation technology.The challenges to robot navigation in complex environment are multifaceted and progressive.In the static environment,the real-time performance of robot navigation,especially the real-time performance of global path planning module,is sensitive to the environment size and requirement of accuracy;For the reason that most path planning algorithms is based on graph search,whose planning time increases sharply with the scale of the working space.So that it is difficult to ensure the real-time performance.In the dynamic environment with many other agents,the overall environmental state becomes much more complex and changeable;Due to the lack of communication in application scenario,it is impossible to correctly predict the environmental state in the future or give countermeasures in advance.And in the interactive environment,the target of robot navigation is no longer limited to the safety performance of the physical level,but also the safety at the spiritual level.As a result,the robot navigation has to be socially compliant.The main research contents and innovations of this paper are as follows:(1)To solve the real-time problem of robot path planning in static environment,and inspired by the path finding method of human-beings,the key crossing points that affect spatial connectivity are obtained and used as the nodes of roadmaps.By using regional segmentation,corner extraction and target recognition,a succinct roadmap that without loss spatial connectivity is found to accelerate the path planning of robot navigation.Compared with the probability completeness of existing sampling methods such as Probabilistic Roadmaps,the spatial discretization strategy based on "gate point" proposed in this paper not only ensures the real-time performance,but also ensures the completeness of the planning algorithm.At the same time,the heuristic function of A* algorithm is also modified in this paper based on reinforcement learning.Compare with A* algorithm,which only describes the end position with a simple Euclidean distance,the proposed method can better capture the obstacle information and avoid the generation of redundant paths.The new planning algorithm based on our new heuristic function achieves a better balance in time and accuracy.In this paper,the real-time problem of robot path planning in static environment is improved from two aspects,a spatial discretization method based on "gate point" is proposed to ensure the completeness,and the heuristic function is modified based on reinforcement learning method to accelerate the search speed.(2)In this paper,a reinforcement learning based algorithm that models the complex interaction between agents using angle pedestrian grid and attention network is proposed,so that the obstacle avoidance and navigation problem in dynamic environment could be solved by the learned strategies.Compared with local grid map encoding or only modeling the interaction between robot and specific agent,the angle pedestrian grid only discretizes the angle information while keep the distance information.Meanwhile,because angle pedestrian grid is much closer to human perception and robot sensor detection principle,the proposed method can better simulate the perception and decision-making state of each agent in the interaction process.Based on the reinforcement learning method,the robot learns appropriate navigation and obstacle avoidance strategy to replace the original local planner module,which avoids the discontinuous control signal caused by the reaction-based methods and the robot freezing problem caused by the trajectory-based methods.In this paper,an optimal obstacle avoidance strategy that imitating human behavior is found by training,the proposed method could deal with obstacle avoidance problem in various scenes and show good real-time performance.(3)In this paper,the global planning and local planning modules are further improved to realize socially compliant navigation in the interaction environment.In order to meet the naturalness requirement of socially compliant navigation,a speed smoothing term is added to the reward function of the local planning module.Meanwhile,the robot’s obstacle avoidance behavior is guaranteed to happen outside the pedestrian’s social space by modifying reward function based on the proxemics model.As a result,the pedestrians could enjoy their private space and the comfort requirements of socially compliant navigation could be achieved.Finally,a path optimizer based on human road rules is designed to add side preference to the global path,so that the robot can abide by the right / left driving principle of human-beings and could better integrate into the human society by meeting the sociability requirements of socially compliant navigation.In this paper,naturalness,comfort and sociability are added to the global planning and local planning of robot navigation respectively,so that the robot navigation and the social rules can be effectively integrated.More specifically,the reward function of local planning is modified to make the robot control signal continuous to meet the naturalness,the reward function is modified based on the proxemics model to make the robot avoid obstacles with additional safety margin to meet the comfort,and the side preference of the global path is ensured by a path optimizer to make the robot abide by the same road rules as human beings and meet the sociability.In this paper,start with the real-time problem of robot navigation in static environment,the global path planning module is optimized by efficient roadmap constructing method and new heuristic function.Then the dynamic environment constraints are introduced and the local planning module is modified based on reinforcement learning method to adapt to the rapid change of environmental state.Finally,social rules are integrated into the global planning and local planning modules of the robot by designing reward function and adding path side preference respectively,and achieved an exploration to the socially compliant navigation.